Gemini Omni Flash's developer API: what's new since I/O

Gemini Omni Flash debuted at Google I/O on May 19 as a consumer-facing conversational video model — live the same day in the Gemini app, YouTube Shorts, YouTube Create, and Google Flow, with a developer API described only as "coming weeks." Six weeks later, on June 30, 2026, that promise landed: Google opened the developer API (gemini-omni-flash-preview), published per-second pricing, and paired the model with a new fast image tier, Nano Banana 2 Lite, built specifically to chain into Omni Flash through the Interactions API.

This piece covers what's new since I/O: the developer-facing specs, the Nano Banana 2 Lite chaining workflow Google is pushing hardest, and where Omni Flash stands with OmniArt today. For the original I/O launch and what Google held back, see our coverage from June 12; for a full spec comparison against Veo 3.1, see our dedicated comparison.

What's new on June 30: the developer API

Gemini Omni Flash (API id gemini-omni-flash-preview) is built for generation and conversational video editing. Confirmed specs for the developer API:

• Pricing: $0.10 per second of video output — the API's own per-second rate, distinct from the consumer subscription pricing (AI Plus, AI Ultra) that covers the May 19 consumer rollout.
• Duration: capped at 10 seconds per generation, unchanged since I/O; Google says longer durations are coming.
• Input: text, image, and video references — though references longer than 3 seconds aren't fully processed, there's no audio-reference upload, and no scene extension in the API yet.
• New availability surfaces: Google AI Studio, the Gemini API, and the Gemini Enterprise Agent Platform — developer and enterprise access, alongside the consumer surfaces (Gemini app, Google Flow) that have been live since I/O.
• Provenance: every output carries SynthID watermarking, verifiable in the Gemini app, Chrome, and Search — unchanged since launch.

Google paired the API opening with a new image tier, Nano Banana 2 Lite (gemini-3.1-flash-lite-image) — about 4 seconds per image, $0.034 for a 1K-resolution output — explicitly designed to chain into Omni Flash through the Interactions API, using previous_interaction_id to carry session history across up to three sequential edits. A range of creative and developer-tool companies, including Astrocade, HubX, Latitude, Stan, Magnific, Agent Opus, Cartwheel, and Flora, tested the pairing ahead of this release.

Current limitations worth knowing before you plan around it

Confirmed gaps as of this API release:

• 10 seconds is the hard cap per generation, unchanged since I/O.
• Video references beyond 3 seconds aren't fully processed.
• No audio-reference upload yet.
• No scene-extension support in the API yet.
• Character consistency across scene changes is an acknowledged weak point, not a solved problem — Google flagged this at I/O and it's still true now.

None of these are disqualifying for a preview-tier API, but they matter for anyone scoping production work: Omni Flash today reads as a fast, short-form, chat-driven iteration tool rather than a broadcast-ready single source of truth for longer narrative video.

How the Nano Banana 2 Lite → Omni Flash chain works

The chaining pattern is the part Google is pushing hardest with this release, and it's demoed through three named workflows:

• "Anywhere" — a photo gets transformed into a new location with Nano Banana 2 Lite, then animated into motion with Omni Flash, all inside one conversational thread.
• "Space Lift" — interior-design stills become cinematic walkthrough video.
• "Omni Product Studio" — a static product photo becomes an e-commerce-ready video.

Mechanically, this runs through the Interactions API: a generation from Nano Banana 2 Lite returns an interaction id, and passing that id as previous_interaction_id into the next call — whether another image edit or an Omni Flash video generation — preserves session history across up to three sequential edits. That's a meaningfully different creative job than plain prompt-to-video: it's prompt to image, refine the image, then animate the refined image, without re-uploading references or re-explaining context at each step. This is the genuinely new piece of this release — it didn't exist at the May 19 I/O launch.

Does Omni Flash clear OmniArt's two bars?

OmniArt's own standard, set when we first covered the pre-I/O leak, is that we add models once they clear two bars: stable public availability and a real creative job the existing lineup doesn't already cover. Omni Flash cleared the first bar back at I/O — it's been publicly available, priced, and documented since May 19.

The second bar is a genuine judgment call, and worth being honest about now that the developer API and Nano Banana 2 Lite pairing exist. The image-to-video chaining workflow — refine a still with a fast image model, then animate it in the same session without re-uploading references — is a legitimate candidate. None of Veo 3.1, Sora 2, V6, Kling 3.0, HappyHorse 1.0, Seedance 2.0, Runway Gen-4.5, Hailuo, or Grok Imagine natively couples to a sibling image model through a session-preserving API the way Omni Flash and Nano Banana 2 Lite now do.

Set against that: a 10-second hard cap, no disclosed resolution figures, weak character consistency across scene changes, and an editing API that doesn't yet support scene extension or audio references. Those are preview-stage limits, not finished-product polish.

Since this article was first published, both halves of that workflow have joined OmniArt: Gemini Omni Flash for standard text- and image-guided video generation, and Nano Banana 2 Lite for fast image generation and reference-driven edits. You can generate or refine a still with Lite, then carry it into Omni Flash or another video model in the workspace. OmniArt does not currently expose Google's single-thread Interactions API controls, so session-preserving conversational edits still require Google's own API surfaces.

We'll keep watching the "real creative job" case as Omni Flash's API moves past preview. For the current lineup and how each model earns its place, see the full video workspace tour.

FAQ

When did Gemini Omni Flash actually launch?

It launched twice, in a sense. Google I/O on May 19, 2026 was the consumer debut — live the same day in the Gemini app, YouTube Shorts, YouTube Create, and Google Flow. The developer API (gemini-omni-flash-preview) followed six weeks later, on June 30, 2026, alongside per-second pricing and the Nano Banana 2 Lite chaining pairing covered in this piece.

How much does Gemini Omni Flash's API cost?

$0.10 per second of video output through the developer API, with generations currently capped at 10 seconds. This is separate from the consumer subscription pricing (AI Plus, AI Ultra) that covers access through the Gemini app and Google Flow.

Is Gemini Omni Flash the same model as Veo?

No. Omni Flash is a distinct, purpose-built video and conversational-editing model with its own API id and pricing, running alongside Veo 3.1 rather than replacing it. For the full spec-by-spec breakdown, see our Omni Flash vs Veo 3.1 comparison.

Can I use Gemini Omni Flash on OmniArt today?

Yes. Gemini Omni Flash is available in OmniArt's video workspace for standard video and reference-image generation. Google's session-preserving conversational editing controls are not exposed in OmniArt's current interface.

What is Nano Banana 2 Lite and how does it relate to Omni Flash?

Nano Banana 2 Lite (gemini-3.1-flash-lite-image) is a new fast, low-cost image tier that launched alongside Omni Flash's developer API — about 4 seconds per image at $0.034 for a 1K-resolution output. It's built to chain directly into Omni Flash through the Interactions API, preserving session history across up to three sequential edits. For how it stacks up against the rest of the Nano Banana 2 family, see Nano Banana 2 Lite vs 2 vs Pro.

Getting started on OmniArt

Open Nano Banana 2 Lite to generate or refine a still, then animate it with Gemini Omni Flash or another model in the video workspace. For a full walkthrough of that image-to-video pattern, see the photo-to-product-video workflow guide.